4 ms·

The problem as I see, at least in the medium term, is that pattern matching, no matter how advanced is still just a simulation of intelligence, not 'real' intel

The problem as I see, at least in the medium term, is that pattern matching, no matter how advanced is still just a simulation of intelligence, not 'real' intelligence. What is missing is the ability to reason, when the facts (data) are inconclusive, missing or wrong. Also any AGI will need a very complete abductive reasoning system, I just don't think we have the tools and understanding yet to build AGI, and certainly not if the next generation of AI practitioners continue to think Python is the way to build an 'intelligence'

mindcrime 5y ago
The problem as I see, at least in the medium term, is that pattern matching, no matter how advanced is still just a simulation of intelligence, not 'real' intelligence. I don't know that we can take that as a given. The "it's just pattern matching" argument is pretty old, but I'm not sure it's ever been shown conclusively that all of human intelligence can't be reduced to some form of pattern matching. Now, in terms of "implement something on a digital computer", so far our approaches to "reasoning" (using symbols and formal logic) seem very, VERY different from our approaches to "pattern matching", and maybe in that domain such a reduction isn't feasible. Or maybe it is. I don't think anybody really knows. My own suspicion (and that's all that it is) is that in "digital computer" domain, we should use formal logic where it makes sense, pattern matching ANN's where it makes sense, and hybridize the systems. But that approach brings in some hard challenges of its own. And FWIW, experts like Geoffrey Hinton have argued against the need for any such hybridization and believe that neural networks (of some form... maybe one we haven't invented yet, who knows?) should be enough for human-level intelligence. I just don't think we have the tools and understanding yet to build AGI I would agree with that. We don't yet have the required understanding (whether or not we have the needed tools is one I'm more 50/50 on), but I do think we'll get there eventually. It's just hard to predict if "eventually" means "tomorrow" or "37,000 years from now".

mindcrime 5y ago

The problem as I see, at least in the medium term, is that pattern matching, no matter how advanced is still just a simulation of intelligence, not 'real' intelligence. I don't know that we can take that as a given. The "it's just pattern matching" argument is pretty old, but I'm not sure it's ever been shown conclusively that all of human intelligence can't be reduced to some form of pattern matching. Now, in terms of "implement something on a digital computer", so far our approaches to "reasoning" (using symbols and formal logic) seem very, VERY different from our approaches to "pattern matching", and maybe in that domain such a reduction isn't feasible. Or maybe it is. I don't think anybody really knows. My own suspicion (and that's all that it is) is that in "digital computer" domain, we should use formal logic where it makes sense, pattern matching ANN's where it makes sense, and hybridize the systems. But that approach brings in some hard challenges of its own. And FWIW, experts like Geoffrey Hinton have argued against the need for any such hybridization and believe that neural networks (of some form... maybe one we haven't invented yet, who knows?) should be enough for human-level intelligence. I just don't think we have the tools and understanding yet to build AGI I would agree with that. We don't yet have the required understanding (whether or not we have the needed tools is one I'm more 50/50 on), but I do think we'll get there eventually. It's just hard to predict if "eventually" means "tomorrow" or "37,000 years from now".